Sample-Efficient Multi-Objective Optimization

125 concepts122 papers3 proposals

When optimizing systems with multiple competing goals—such as maximizing a machine learning model's accuracy while minimizing its fairness bias, or maximizing a vehicle's tracking precision while minimizing its energy consumption—the traditional instinct is to collapse these goals into a single number. This technique, known as scalarization or the weighted-sum approach, assigns a fixed weight to each objective to create a single-objective problem.

However, as noted by Conor F. Hayes et al. in A practical guide to multi-objective reinforcement learning and planning, scalarization forces a premature compromise. It requires running an entire optimization pipeline repeatedly with different weight combinations just to map out the trade-off space. This is highly sample-inefficient and completely blind to non-convex regions of the true Pareto front—the set of optimal solutions where no single objective can be improved without degrading another.

Direct multi-objective optimization avoids this trap by treating objectives as vectors rather than scalars, finding a diverse menu of optimal trade-offs in a single optimization pass.

Why Vector-Valued Search Works

The core reason direct multi-objective optimization is more sample-efficient than repeated scalarization comes down to information sharing. In a scalarized setup, an evaluation tells the algorithm only how well a solution performed under one specific weighting. If you change the weights, the algorithm must largely start over. In contrast, vector-valued optimization evaluates the potential improvement across all objectives simultaneously.

For instance, in Bayesian Optimization (BO), an acquisition function like the Expected Improvement Matrix (EIM) can identify the non-dominated Pareto-optimal set in a single run. As demonstrated by Ali Gharib et al. in Multi-Objective Optimization of a Path-following MPC for Vehicle Guidance: A Bayesian Optimization Approach, this direct approach achieves comparable or superior solution quality with nearly an order of magnitude fewer evaluations than traditional scalarization.

When navigating complex physical constraints where each evaluation requires an expensive simulation, this efficiency is critical. For example, Alexander Thebelt et al. showed in Multi-Objective Constrained Optimization for Energy Applications via Tree Ensembles that replacing standard Gaussian Processes with tree ensembles and optimizing the acquisition function globally allows algorithms to find high-quality Pareto frontiers far more efficiently than state-of-the-art evolutionary algorithms like NSGA-II.

Scaling with Local Structure and Gradients

Historically, direct multi-objective optimization relied heavily on black-box genetic algorithms, which struggle to scale as search spaces grow larger or more complex. Recent breakthroughs have introduced local structure and gradient acceleration to overcome this bottleneck.

To handle high-dimensional engineering tasks, Samuel Daulton et al. developed regionalized approaches in Multi-Objective Bayesian Optimization over High-Dimensional Search Spaces, coordinating local optimization regions to maintain sample efficiency. Furthermore, researchers are no longer restricted to derivative-free search. By reformulating multi-objective learning into first-order penalty methods, Lisha Chen et al. introduced FOOPS in Efficient First-Order Optimization on the Pareto Set for Multi-Objective Learning under Preference Guidance, achieving superior hypervolume coverage without the massive computational overhead of second-order derivatives.

To accelerate convergence even further, Konstantin Sonntag et al. demonstrated in Fast Multiobjective Gradient Methods with Nesterov Acceleration via Inertial Gradient-Like Systems that treating optimization as a continuous-time inertial dynamical system can reduce the required iteration count by roughly 75% while preserving a diverse spread of solutions across both convex and non-convex landscapes.

Conquering the Many-Objective Frontier

When an optimization problem expands from two or three objectives to dozens (a subfield known as many-objective optimization), a unique mathematical challenge arises: "dominance resistance." In high-dimensional objective spaces, almost every random solution becomes non-dominated relative to the others. As a result, standard Pareto-sorting mechanisms stall because they can no longer distinguish between a high-quality trade-off and a mediocre solution that happens to score well on just one obscure metric.

To counteract this, modern frameworks use decomposition and structured geometric partitioning. For example, Bilel Derbel et al. proposed a set-oriented approach in A set-oriented MOEA/D to explicitly map sub-problems to solutions, maintaining steady convergence as objectives scale. To push this further, Vikas Palakonda et al. utilized radial grids and external archives in External archive guided radial-grid multi objective differential evolution to successfully navigate up to 15 conflicting goals simultaneously.

By combining decision-space dimension reduction with diversity-focused sampling, the authors of Improved Regularity Model-based EDA for Many-objective Optimization (Yanan Sun et al.) demonstrated that algorithms can achieve convergence up to twice as fast as standard evolutionary methods, even when dealing with inverted or highly complex Pareto front shapes that typically cause other multi-objective algorithms to fail.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_multi_objective_optimization_performance_dc3b9230,
  title = {multi-objective optimization performance},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-objective-optimization-performance/rdr_dc3b923074f5421b8f26eea62e4c10f3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_multi_objective_optimization_performance_dc3b9230_v1,
  title = {multi-objective optimization performance},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-objective-optimization-performance/rdr_dc3b923074f5421b8f26eea62e4c10f3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake